Saffron authenticity is important for the saffron industry, consumers, food industry, and regulatory agencies. Herein we describe a combo of two novel methods to distinguish genuine saffron from fake in a user-friendly manner and without sophisticated instruments. A smartphone coupled with Foldscope was used to visualize characteristic features and distinguish "genuine" saffron from "fake." Furthermore, destaining and staining agents were used to study the staining patterns. Toluidine blue staining pattern was distinct and easier to use as it stained the papillae and the margins deep purple, while its stain is lighter yellowish green toward the central axis. Further to automate the process, we tested and compared different machine learning-based classification approaches for performing the automated saffron classification into genuine or fake. We demonstrated that the deep learning-based models are efficient in learning the morphological features and classifying samples as either fake or genuine, making it much easier for end-users. This approach performed much better than conventional machine learning approaches (random forest and SVM), and the model achieved an accuracy of 99.5% and a precision of 99.3% on the test dataset. The process has increased the robustness and reliability of authenticating saffron samples. This is the first study that describes a customer-centric frugal science-based approach to creating an automated app to detect adulteration. Furthermore, a survey was conducted to assess saffron adulteration and quality. It revealed that only 40% of samples belonged to ISO Category I, while the average adulteration percentage in the remaining samples was 36.25%. After discarding the adulterants from crude samples, their quality parameters improved significantly, elevating these from ISO category III to Category II. Conversely, it also means that Categories II and III saffron are more prone to and favored for adulteration by fraudsters.
Keywords: Crocus sativus; Foldscope; adulteration; deep learning; fraud; image processing; machine learning; microscopy.
Copyright © 2022 Husaini, Haq, Shabir, Wani and Dedmari.